On Mon, 18 Jun 2012 23:56:18 -0300 Henrique de Moraes Holschuh <h...@debian.org> wrote:
> On Mon, 18 Jun 2012, Stephen Hemminger wrote: > > I will be happy to get involved; upstream irqbalance is on my list > > of things that are broken and needs to be fixed. Will try and get a hold > > Could you elaborate on this? > > > of past contributors to figure out what is happening. I know there was > > talk of fixing it and integrating Holger's irqd logic. > > Don't cache topology and sibling relationship also play a role on maximum > PPS throughput NIC interrupt routing when assigning interrupts to proper > multiqueue NICs? Especially when Intel DCA is enabled? > > Although trying to route the packet while it is still cache-hot might as > well be a pipe dream... > I haven't looked in detail at current version. But the older versions had dynamic policy and depended on making assumptions about which device had which irq (by looking at /proc/interrupts), and polling for network activity. What it was trying to do might have been a good idea in old i386 days but was getting horribly confused with all the types of multiqueue drivers etc. The ideal program would be state driven, handle power management (use less cores when not busy), and be dynamically user configurable, and solve world hunger... -- To UNSUBSCRIBE, email to debian-devel-requ...@lists.debian.org with a subject of "unsubscribe". Trouble? Contact listmas...@lists.debian.org Archive: http://lists.debian.org/20120619153455.4fa54...@s6510.linuxnetplumber.net
